Article Text

Download PDFPDF
Short report
Can brain signals and anatomy refine contact choice for deep brain stimulation in Parkinson’s disease?
  1. San San Xu1,2,3,
  2. Wee-Lih Lee1,
  3. Thushara Perera1,3,
  4. Nicholas C Sinclair1,3,
  5. Kristian J Bulluss1,4,5,6,
  6. Hugh J McDermott1,3,
  7. Wesley Thevathasan1,2,7,8
  1. 1Bionics Institute, East Melbourne, Victoria, Australia
  2. 2Department of Neurology, Austin Hospital, Heidelberg, Victoria, Australia
  3. 3Medical Bionics Department, The University of Melbourne, Melbourne, Victoria, Australia
  4. 4Department of Neurosurgery, St Vincent's Hospital, Fitzroy, Victoria, Australia
  5. 5Department of Neurosurgery, Austin Hospital, Heidelberg, Victoria, Australia
  6. 6Department of Surgery, The University of Melbourne, Parkville, Victoria, Australia
  7. 7Department of Neurology, The Royal Melbourne Hospital, Parkville, Victoria, Australia
  8. 8Department of Medicine, The University of Melbourne, Parkville, Victoria, Australia
  1. Correspondence to Dr Wesley Thevathasan, Bionics Institute, East Melbourne, VIC 3050, Australia; Wesley.Thevathasan{at}mh.org.au

Abstract

Introduction Selecting the ideal contact to apply subthalamic nucleus deep brain stimulation (STN-DBS) in Parkinson’s disease is time-consuming and reliant on clinical expertise. The aim of this cohort study was to assess whether neuronal signals (beta oscillations and evoked resonant neural activity (ERNA)), and the anatomical location of electrodes, can predict the contacts selected by long-term, expert-clinician programming of STN-DBS.

Methods We evaluated 92 hemispheres of 47 patients with Parkinson’s disease receiving chronic monopolar and bipolar STN-DBS. At each contact, beta oscillations and ERNA were recorded intraoperatively, and anatomical locations were assessed. How these factors, alone and in combination, predicted the contacts clinically selected for chronic deep brain stimulation at 6 months postoperatively was evaluated using a simple-ranking method and machine learning algorithms.

Results The probability that each factor individually predicted the clinician-chosen contact was as follows: ERNA 80%, anatomy 67%, beta oscillations 50%. ERNA performed significantly better than anatomy and beta oscillations. Combining neuronal signal and anatomical data did not improve predictive performance.

Conclusion This work supports the development of probability-based algorithms using neuronal signals and anatomical data to assist programming of deep brain stimulation.

  • EVOKED POTENTIALS
  • PARKINSON'S DISEASE
  • ELECTRICAL STIMULATION
  • NEUROSURGERY
  • NEUROPHYSIOLOGY

Statistics from Altmetric.com

Introduction

The efficacy of subthalamic nucleus (STN) deep brain stimulation (DBS) for patients with Parkinson’s disease (PD) relies on delivering stimulation to the ideal location.1 After implantation, stimulation location is determined by the selection of contact(s) on each lead by clinicians. Identifying the ideal contact is an arduous, time-consuming and ultimately heuristic process dependent on expertise. A potential solution is to develop aids that use objective data including anatomy2 3 or neuronal signals recorded from DBS leads to guide, and eventually automate, contact selection. Beta oscillations4 5 have been explored for this purpose, and more recently, the novel signal evoked resonant neural activity (ERNA).6–10 Here, in a cohort of 92 hemispheres in 47 patients with PD receiving STN-DBS, we assessed how anatomy and neuronal signals could be used individually, or in combination, to predict the contacts chosen by expert-clinician programming at 6 months postoperatively.

Methods

Cohort selection

We accessed data of 100 hemispheres from 50 consecutive patients implanted with quadripolar leads (Medtronic 3387) between 8 February 2016 and 21 March 2019. Only hemispheres receiving monopolar and bipolar DBS at 6 months postoperatively were included. Thus, the final study cohort comprised of 92 hemispheres (75 monopolar, 17 bipolar) in 47 patients.

Patients were implanted by a single DBS team in Melbourne, Australia, using surgical methods described previously.6 10 Postoperatively, programming was performed by two experienced DBS neurologists (WT, SSX) from a high-volume DBS service (>50 denovo implantations annually). Clinicians were blinded to the neuronal signal recordings, but routinely visualised electrode location by manually fusing the preoperative MRI and postoperative CT (StealthStation S7, Medtronic, Dublin, Ireland) to assist programming. Monopolar configuration was selected initially, changing to bipolar or interleaved configuration if clinically indicated. Dorsal interleaved configuration (in five hemispheres) was employed to minimise dyskinesias and classified as monopolar, ignoring the dorsal contact. Programming information was accessed from DBS programmer reports. Neuronal signal and anatomical data have been previously reported in 13 patients.10 11 The study was approved by the St Vincent’s Public (HREC-D 071/14), St Vincent’s Private (R0236-15), Austin (SSA/15/Austin/266) and Cabrini (02-15-02-16) Human Research Ethics Committees.

Neuronal signal recording and analysis

Immediately after lead implantation, neuronal activity was recorded at each contact in a monopolar configuration and rereferenced to an average of the four unstimulated contacts in the contralateral hemisphere. Neuronal signal analysis was performed using MATLAB R2017a (Mathworks, Massachusetts, USA).

Burst stimulation was sequentially applied to each contact at a rate of one burst per second for 10 s. Each burst contained 10 symmetric biphasic pulses delivered at 3.375 mA, 60 μs and 130 Hz.6 ERNA was recorded from the stimulating contact. ERNA amplitude was measured by averaging the root mean square amplitude from 3.5 to 20 μs after the onset of the last pulse of each burst and squared to calculate power.6 Beta power was estimated over the 13–30 Hz frequency band from Blackman-Harris windowed epochs of 1 s processed using a short-time Fourier transform.7 For machine learning analysis, neuronal signals were normalised by dividing signal power recorded at a given contact by the sum of the signal power across the four contacts in each hemisphere.

Contact anatomical mapping

The preoperative 3 Tesla MRI was reoriented to align the anterior and posterior commissures using an automated utility (NITRC, acpcdetect, V.2.0). The MRI and postoperative CT were coregistered (BRAINSFit V.4.11) and normalised into the Montreal Neurological Institute (MNI) 152 2009b space using Advanced Normalisation Tools SyN. The DBS leads were localised on CT using the PaCER toolbox and visually confirmed (figure 1A). Resulting contact coordinates were transformed into MNI space. The Euclidean distance between each contact and an ideal anatomical location to apply STN-DBS, nominated according to MNI coordinates provided by Horn et al,12 was calculated using a custom Python script. The DISTAL atlas was used to segment the STN and surrounding structures for 3D visualisation.

Figure 1

(A) Placement of deep brain stimulation (DBS) electrodes visualised in the Montreal Neurological Institute (MNI) space. Blue = globus pallidus externa. Yellow = globus pallidus interna. Green = subthalamic nucleus. (B) Confusion matrix depicting the mean predictive value (MPV) at each factor ranking using the simple-ranking model (of evoked resonant neural activity (ERNA) power, beta oscillation power and anatomy) and random forest model (of normalised ERNA power, normalised beta power and anatomy). All models were validated using 10-fold nested cross-validation, repeated 10 times. For example, in the simple-ranking model of ERNA power, the MPV of ranking 1 (top row, left corner) was 80% and ranking 2 (top row, second square from the left) was 17%. Darker squares denote a higher MPV. (C) The MPV of first-ranked contacts according to ERNA power, beta oscillation power and anatomical location using the simple-ranking method and random forest classifier. Bars represent standard deviation. P values in figures are adjusted for multiple comparisons (Nemenyi test).

Statistical analysis

The variation in neuronal signal power between adjacent contacts on the DBS lead was evaluated using a mixed-effects model, with Bonferroni-Holm method employed for post hoc analysis (online supplemental material). Two methods were used to assess how anatomy and neuronal signals could be used individually, or in combination, to predict the contacts chosen by clinicians at 6 months postoperatively. First, in a simple-ranking method, contacts were ranked within each hemisphere according to ERNA power, beta power and Euclidean distance from the ideal anatomical location to apply STN-DBS. These rankings were compared with the contacts selected for chronic DBS at 6 months (the cathode (−) in monopolar or bipolar configuration). Second, machine learning algorithms (logistic regression (LR), support vector machine (SVM), random forest (RF)) incorporating normalised ERNA, normalised beta and anatomical data were applied to explore whether combining factors could improve the probability of predicting the clinician-chosen contact (online supplemental material). The two methods were validated using 10-fold nested cross-validation (CV), repeated 10 times. The probability that the model predicted the clinician-chosen contact (predictive value) was calculated as the mean performance of the nested CV folds. Differences between models were evaluated using the Friedman rank sum test and Nemenyi test. Machine learning algorithms and confusion matrices were built using Scikit-Learn V.0.23.2 on Python V.3.8.5 (Delaware, USA). All other statistical analyses were performed using R Project V.4.1.2 (Vienna, Austria).

Supplemental material

Data availability

Anonymised data are available on request for the purpose of replicating procedures and results, subject to an embargo of 24 months from the date of publication.

Results

Patient characteristics are outlined in table 1.

Table 1

Participant characteristics

Contacts ranked by ERNA power, beta power and anatomy

Log-transformed ERNA power (r2=0.84, Akaike Information Criterion (AIC) 344.6, p<0.001) and beta power (r2=0.86, AIC −356.6, p<0.001) varied significantly among adjacent contacts (online supplemental figure 5).

The first-ranked contacts according to ERNA power, beta power and anatomy matched the clinician-chosen contacts in 74/92, 45/92 and 62/92 hemispheres, respectively. In bipolar hemispheres, when the first-ranked ERNA and anatomy contacts did not match the clinician-chosen cathode (−) (ERNA – 4/17 hemispheres; anatomy – 8/17 hemispheres), they always corresponded with the clinician-chosen anode (+). In the simple-ranking model, the probability that the first-ranked contact according to ERNA power, beta power and anatomy was selected by the clinician was 80.3% (SD 14.2), 49.9% (SD 16.5) and 66.8% (SD 16.5), respectively (figure 1B).

Single-factor and multifactor model performance

There was a difference between ERNA, beta oscillations and anatomy models built using the simple-ranking method (Friedman χ2=111.4, p<0.001). ERNA was more predictive than beta oscillations and anatomy (Nemenyi test, both p<0.001). Anatomy was more predictive than beta oscillations (Nemenyi test, p<0.001) (figure 1C).

Clinician-chosen contacts predicted by anatomy usually also corresponded with the first-ranked contact according to ERNA (56/62 hemispheres). Clinician-chosen contacts predicted by ERNA corresponded with either the first-ranked (56/74 hemispheres) or second-ranked (17/74 hemispheres) contact according to anatomy. There was a lower correspondence rate between the clinician-chosen contacts predicted by beta and those predicted by ERNA and anatomy (online supplemental figure 6).

Of the three classification algorithms (LR, SVM and RF) explored to evaluate the model performance of combinations of factors, the RF classifier performed best and was selected for subsequent analysis (online supplemental table 1). The probability that ERNA data predicted the clinician-chosen contact was 80.7%. This did not change significantly with inclusion of anatomy and/or beta oscillation information (Friedman χ2=6.1, p=0.10) (figure 1C).

Discussion

In 92 hemispheres of 47 patients implanted with STN-DBS for PD, we found that anatomy and neuronal signals predicted the contacts selected by expert-clinician programming. The predictive value of each factor was as follows: ERNA 80%, beta oscillations 50%, anatomy 67%. ERNA was significantly more predictive than beta oscillations and anatomy and the addition of beta oscillations and anatomical information did not improve model performance.

Here, we employed a real-world outcome of the contacts selected by expert clinicians as the ‘gold standard’. However, the DBS programming heuristic can be complex, time-consuming and prone to error. Nonetheless, the programming clinicians were experienced and patients achieved an 89.1% reduction in dopaminergic medication doses with DBS. An alternative approach would be to systematically assess the acute motor impact of DBS at every contact in experimental sessions.4 10 However, acute studies are liable to confounds such as incomplete washout and missing DBS effects that evolve over longer timescales.

The error prone nature of DBS programming highlights the need for objective data to guide contact choice. Nonetheless, no factor alone predicted the clinician-chosen contact in all hemispheres. For example, 80% of contacts ranked first by ERNA were clinically selected. It is unknown whether the other 20% of contacts ranked first by ERNA would have achieved equivalent or better outcomes than those selected by the clinicians. Only 67% of contacts ranked first by anatomy were clinically selected for chronic DBS (even though clinicians used lead location to assist programming). Perhaps the performance of structural anatomy was confounded by the normalisation of images into the MNI space to allow standardised assessment of the ‘ideal’ location across hemispheres. However, such methods are a well-validated approach for such a data set.13 Furthermore, structural anatomy may not always indicate the ideal functional location for neuromodulation. Supporting this idea, we found substantial discordance between the rankings of ERNA versus anatomy at the clinician-chosen contacts. In future, refinements to signal processing or the use of native imaging or advanced imaging techniques14 15 could better predict the ideal contact to apply DBS in individual patients. For example, neuronal signals and structural imaging may not inform on aberrant fibre tracts15 that can cause side effects.

This study assessed patients implanted with quadripolar DBS leads with 1.5 mm spacing between adjacent contacts. Newer devices can employ leads with smaller spacing between contacts, including those with directional and multiple independent current control.16 These capabilities carry a much greater programming complexity. Clinician aids that employ lead localisation to assist programming have been developed.1–3 12 Such methods require expert image acquisition and analysis, and accounting for the orientation of directional electrodes is challenging.17 Beta oscillations have also been explored to guide programming in directional leads.4 5 However, in this study, we found that ERNA was more predictive of the clinician-chosen contact compared with both anatomy and beta oscillations. While the performance of ERNA in narrow spaced electrodes needs to be assessed, we found a promising spatial resolution with a large power gradient between the first-ranked and remaining contacts. Our study raises the potential that ERNA, captured by intraoperative recording systems or even next-generation implantable pulse generators, could be developed to assist programming.

Ethics statements

Patient consent for publication

Ethics approval

The study was approved by the St Vincent’s Public (HREC-D 071/14), St Vincent’s Private (R0236-15), Austin (SSA/15/Austin/266) and Cabrini (02-15-02-16) Human Research Ethics Committees. Participants gave informed consent to participate in the study before taking part.

References

Footnotes

  • Contributors SSX was involved with the conception, organisation and execution of the research project, the statistical analysis and the writing of the first draft of the manuscript. W-LL was involved in the organisation and execution of the research project and the statistical analysis and drafting, review and critique of the manuscript. TP was involved in the organisation and execution of the research project and review and critique of the manuscript. NCS was involved in the conception, organisation and execution of the research project and the review and critique of the manuscript. KJB was involved in the organisation and execution of the research project and the review and critique of the manuscript. HJM was involved in the organisation and execution of the research project and the review and critique of the manuscript. WT was involved in the conception, organisation and execution of the research project and the review and critique of the manuscript.

  • Funding This study was funded through the National Health and Medical Research Council (Development grant number 1177815, project grant number 1 103 238 (Bionics Institute), post graduate scholarship number 1 133 295 (SS.X)), St Vincent’s Hospital Research Endowment Fund and Colonial Foundation. WT is also supported through Lions International. All authors affiliated with the Bionics Institute acknowledge the support it receives from the Victorian Government through its operational infrastructure programme.

  • Competing interests SSX holds options in DBS Technologies Pty Ltd. W-LL has no relevant financial disclosures. TP receives consulting fees and holds options in DBS Technologies Pty Ltd and is a named inventor on related patents, which are assigned to DBS Technologies Pty Ltd. NCS is a named inventor on related patents, which are assigned to DBS Technologies Pty Ltd. KJB, HJM and WT are co-founders and hold shares and options in DBS Technologies Pty Ltd which plans to commercialise the use of neuronal signals to improve DBS. KJB, HJM and WT are also named inventors on related patents, which are assigned to DBS Technologies Pty Ltd.

  • Provenance and peer review Not commissioned; externally peer reviewed.

  • Supplemental material This content has been supplied by the author(s). It has not been vetted by BMJ Publishing Group Limited (BMJ) and may not have been peer-reviewed. Any opinions or recommendations discussed are solely those of the author(s) and are not endorsed by BMJ. BMJ disclaims all liability and responsibility arising from any reliance placed on the content. Where the content includes any translated material, BMJ does not warrant the accuracy and reliability of the translations (including but not limited to local regulations, clinical guidelines, terminology, drug names and drug dosages), and is not responsible for any error and/or omissions arising from translation and adaptation or otherwise.

Request Permissions

If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.